|
Abstract:
|
Randomization is the 'gold standard' to estimate casual effects, but chance imbalance exists in covariate distribution among treatment groups. To address this issue, we propose a new covariate-adaptive design to improve the covariate balance. We specify an explicit definition of imbalance and control it by assigning units sequentially and adaptively. If covariates vary in importance, we partition them into tiers to ensure that important covariates have better balance. With a large number of covariates or a large sample size, our method has substantial advantages over traditional methods in terms of the covariate balance and computational time, and as such becomes an ideal technique in the era of big data. More crucially, our method attains the optimal covariate balance, in the sense that the estimated average treatment effect under our method attains its minimum variance asymptotically. All the above mentioned advantages of our method are further evidenced by extensive simulation studies.
|
